Honor `tools:` frontmatter when an agent is invoked via `@mention` — strip `Task` only when the agent did not declare it

Summary

When a Claude Code agent is invoked via @mention from inside an already-active session, Claude Code wraps the invocation as Task(subagent_type=<agent>, ...). The runtime then strips Task from the agent's available toolset as an anti-recursion guard — even when the agent's tools: frontmatter explicitly declares Task.

This breaks hub-style agents that exist specifically to orchestrate other subagents. The user's mental model when typing @th:orchestrator is "the orchestrator agent runs and dispatches subagents". The actual behavior is "the orchestrator runs, detects it has no Task, and produces a verbose handoff for top-level Claude to execute".

The workaround works (handoff pattern is well-documented and self-healing) but introduces:

• A round-trip turn — the orchestrator hands off, then top-level Claude executes the dispatch.
• ~3k tokens of handoff playbook per invocation (mitigated to ~300 tokens by Option 4 fix in our harness; see https://github.com/valianx/team-harness/pull/15).
• UX inconsistency between @<orchestrator-style> (anti-pattern) and @<leaf-style> (works as expected).

Proposed change

When Claude Code dispatches an agent via Task(subagent_type=X, ...) (either user-triggered via @mention or orchestrator-triggered via inline Task call), honor the agent's declared tools: frontmatter:

• If the agent's frontmatter lists Task → permit Task in the nested invocation.
• If the agent's frontmatter omits Task → strip as today.

The anti-recursion guard becomes opt-in via the frontmatter declaration, which is the existing capability mechanism. Agents that don't declare Task see no behavior change. Hub agents that explicitly declare Task get to do their job.

Why this is safe

1. tools: frontmatter is already the source of truth for what an agent can call. It's the declarative contract. Today's behavior contradicts that contract (silently strips a declared tool).
2. Anti-recursion concern is reasonable but already mitigated: a hub agent calling itself (Task(subagent_type=orchestrator) from inside orchestrator) can be blocked specifically without blocking Task to call OTHER subagents.
3. Opt-in via existing mechanism: no new config keys, no new flags. The frontmatter that already exists tells the runtime what the agent can call.
4. Backward compatible: agents that don't declare Task (the common case) are unaffected.

Concrete example

Our harness team-harness (https://github.com/valianx/team-harness) has an orchestrator agent with:

---
name: orchestrator
tools: [Read, Edit, Write, Bash, Glob, Grep, Task, ...]
---

When invoked from top-level (e.g., user opens a session and Task is dispatched once to orchestrator), Task is available and the orchestrator can call Task(subagent_type=architect, ...), etc.

When invoked via @th:orchestrator mention from inside an active session, Task is stripped despite being declared in tools:. The orchestrator detects this via a boot probe and emits a dispatch handoff. Top-level Claude reads the handoff and takes over the orchestration role.

Alternatives considered

1. Allow recursion entirely (no anti-recursion guard): rejected — too easy to get into infinite loops.
2. Recursion depth limit (allow Task until depth N, then strip): workable but more complex than the proposed change.
3. New nested_tools: frontmatter field: rejected — duplicates tools:.
4. Status quo + workaround: what we do today. Works but is the source of this request.

Impact

• Affected users: anyone who has built hub-style agents (orchestrator, planner, multi-stage pipeline agents).
• UX improvement: @<hub-agent> works as the name implies.
• Token reduction: eliminates handoff round-trip when the hub agent is invoked via @mention.

Workaround we ship today

Documented at https://github.com/valianx/team-harness/blob/main/agents/orchestrator.md (Dispatch-blocked exit section) and https://github.com/valianx/team-harness/blob/main/CLAUDE.md (§14 Universal rule — auto-takeover on blocked-no-dispatch). PR https://github.com/valianx/team-harness/pull/15 reduces the handoff token cost from ~3k to ~300 via a structured JSON payload, but the round-trip itself remains until this upstream change lands.

Environment

• Claude Code CLI version: any recent (issue is consistent across releases).
• Reproducibility: deterministic via @th:orchestrator mention on any agent that declares Task.

Related

• Our harness issue tracking this from the consumer side: https://github.com/valianx/team-harness/issues/14
• Mitigation PR (Option 4): https://github.com/valianx/team-harness/pull/15